- 1. Medical Research Council, Memorandum No. 45: Aids to the examination of the peripheral nervous system, Crown, London, 1976.
- 2. E F Delagi and A O Perotto: Anatomic Guide for the Electromyographer, C C Thomas, Springfield, Ill. 1980.
- 3. R K Sethi and L L Thompson: The Electromyographer's Handbook, Little, Brown and company, Boston, 1989.
- 4. J Kimura: Electrodiagnosis in Diseases of Nerve and Muscle, F A Davis, Philadelphia, 1989.
- 5. J A Liveson: Peripheral Neurology, FA Davis Company, Philadelphia, 1991.
- 6. R D Adams and M Victor: Principles of Neurology, McGraw-Hill, New York, 1993.
- 7. A O Perotto: Anatomic Guide for the Electromyographer: The Limbs and Trunk, C C Thomas, Springfield, Ill., 1994.
- 8. L P Rowland: Merritt's Textbook of Neurology, Williams and Wilkins, Baltimore, 1995.
- 9. R J Joynt and R C Griggs: Clinical Neurology, Lippincott Williams and Wilkins, Philadelphia, 1998.

DETAILED DESCRIPTION

The diagnosis of focal nerve damage is based in part on the evaluation of muscles, usually those that are clinically related to the patient's complaints. This examination can be a physical examination and/or electromyographic examination of said muscles. This examination will generate and record abnormality levels for each one of the muscles examined, usually in the range of 0 to 3, where zero stands for normal muscle and 3 stands for maximally damaged muscle. This range can be expanded.

Note: By design the EMG ASSISTANT interprets muscle data that was generated by the clinician; it does not interpret raw EMG signals. Also, it is designed to localize focal nerve damage and it is not diagnostic in polyneuropathy and not in myopathy.

The EMG ASSISTANT analysis of the patient's clinical findings is based onFIGS. 1 & 2. These Figures represent nerve-to-muscle connections in matrix formats and are the condensation of known anatomical and pathological data (references 1-9).

In the present version, the upper extremity (UES) data is condensed and limited to 27 muscles to be examined and 26 possible locations where upper extremity nerves may be damaged (Table 1). This system can be extended and applied to much bigger matrices; more muscles and more nerves. The locations that were chosen for the present version are those that are presently generally accepted nerve damage locations in clinical neurophysiology. Innervation-aberrations are not accounted for and have not been automated; they are left for the clinician to consider.

The lower extremity (LES) data is condensed and limited to 27 muscles that can be examined and 17 locations where lower extremity nerves may be damaged (Table 2). These locations are also the accepted nerve damage locations in clinical neurophysiology and can be extended to check more muscles and more nerves. Innervation-aberrations are not accounted for and have not been automated; they are left for the clinician to consider.

Steps in the Process Relating to the Upper Extremities (UES):

Step 1 (FIG. 3, A): This step does not involve patient's data. The program begins by generating all the possible one-to-many virtually damaged nerve locations in the UES nerve-sets (1-to-26 possible locations, n1 to n26 inFIG. 1). This process will generate 2²⁶=67,108,864 nerve-sets. Each nerve set contains the names of 26 nerves—some of which are damaged, some of which are not. Each set is unique. Zero stands for a nerve location that is unaffected, i.e. normal, and One stands for a damaged nerve location within a set. As this in not empirical data and there is no information on the severity of the damage, it is restricted to 0s and 1s.

Step 2 (FIG. 3, B): This step does not involve patient's data. Each one of the 2²⁶unique nerve-sets above is translated by the program usingFIG. 1 into the theoretically expected muscle-sets (m1 to m27 inFIG. 1). This process will generate 2²⁶=67,108,864 sets of muscles. Within each set some of the muscles are normal, and some of the muscles are damaged. These 2²⁶muscle-sets may not be unique as several combinations of nerve-sets may theoretically bring about the same muscle-set. Zero stands for muscles that are not affected by damaged nerves in the corresponding nerve-sets, i.e. normal, and One stands for abnormal muscles within the 27 muscles of each set. As this in not empirical data, it is restricted to 0s and 1s.

Step 3 (FIG. 3, C): The clinician examines the patient and records the actual muscle-set, P1-to-P27 (right upper line in the Figure), specific to this patient, and inputs it into the computer using the input dialog. At this stage in the explanation of the EMG ASSISTANT we assume that the clinician grades the patient's muscles in a restricted range of only 0s and 1s; see below for an expanded explanation.

Step 4 (FIG. 3, D): The program compares the actual patient's muscle-set (P1-to-P27) to everyone of the previously generated 2²⁶muscle-sets (m1-to-m27). When identity is found—there might be more than one identity—the nerve-set that generated this identity is found as marked by the broken arrow (see more below): This is one of the possible diagnoses. Then, from all the possible diagnoses the program selects the nerve-set(s) with the smallest number of damaged nerves and reports it (them) as the final diagnosis.

The above statedSteps 3 & 4 are shown in order to simplify the reader's learning process. However, almost always the clinician will grade the muscles in a wide range of abnormalities, such a 0, 1, 2 and 3, and sometimes even into wider ranges. Therefore, the comparisons as done inFIG. 3, D do not amount to identities, since the comparisons of 0s and 1s to 2s and 3s cannot amount to identities. A diagnosis based on the clinical severity of the data is likely to be more reliable than that resulted from the restricted range of normal/abnormal (0s and 1s). The actual comparisons are based on either the sums of absolute-value differences or the sum-of-squared differences between sets of P1-to-P27 and m1-to-m27 following certain rules. The clinician can set the program to whichever rule he/she is comfortable with.

Step 5: See below under “OUTPUT.”

Rules of Computation, Version 1 (FIG.4, A)

Each one of the 2²⁶computer generated virtual muscle-sets (m1-to-m27) is compared with the clinician's empirical muscle-set (P1-to-P27), one corresponding muscle at a time (m1 with P1 through m27 with P27, generalized to Px and mx), and for each pair (Px, mx) the 27 separate differences are added up to that set's sum-total:

- a. If Px=−1, i.e. this muscle was not examined by the clinician and will not be counted.
- b. If Px=0 and mx=1: The difference is set to NV (defaults to 2).
- c. If Px>0 and mx=0: The difference if set to Px.
- d. All other possible comparisons are not considered.
- e. All comparisons above are added to give the sum-total for each of the computer's generated 2²⁶muscle-sets.

EXAMPLE


Px:	−1 −1 0 0 1 1 2 2 3 3 −1 −1 0 0 1 1 2 2 3 3 −1 −1 0 0 1 1 2
mx:	0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0
Diff:	0 0 0 2 1 0 2 0 3 0 0 0 0 2 1 0 2 0 3 0 0 0 0 2 1 0 2

Sum − total = 21

- f. This method's sum-total values can range from 27*3=81 (when each Px=3 and each mx=0) to 27*0=0 (when each Px>0 and each mx=1).
- g. The muscle-sets with the smallest sum-totals are considered best fit and are converted back to the nerve-sets that could have generated them. There may be more than one such nerve-set).
- h. From the latter, the one or the few nerve-sets with the smallest number of damaged nerves are output as the diagnoses.

The rules of comparison for sum of squared-differences are identical to the above only that the sum is the sum of the squared-differences. In the example above with squared-differences the sum-total=45.

The logic behind these rules is as follows: As mx can be either 0 or 1, where 1 indicates abnormal muscle and 0 indicates normal muscle, supposedly we cannot measure the difference between any empirical grade of abnormality higher than 1 and that of virtual 1; however the differences when Px>0 and mx=0 is measurable by the certainty of the examiner that that muscle is abnormal and its abnormality equals Px. Also, the difference when Px=0 and mx=1 is measurable and is the level of the clinician's confidence in his statement that the muscle (Px) is normal when compared to the computer generated value of 1. In this software that confidence is called NV (Normal Value) and defaults to 2.

Rules of Computation, Version 2 (FIG.4, B)

Each one of the computer generated 2²⁶virtual muscle-sets is compared with the clinician's actual muscle-set, one corresponding muscle at a time (m1 with P1 through m27 with P27, generalized as Px and mx) and for each compared set-pair the 27 separate differences are added-up to that set's sum-total:

- a. If Px=−1, this muscle was not examined and will not be counted.
- b. If Px=0 and mx=0: The difference is set to −NV (NV defaults to 2).
- c. If Px=0 and mx=1: The difference is set to +NV
- d. If Px>0 and mx=0: The difference is set to Px.
- e. If Px>0 and mx=1: The difference is set to −Pk.
- f. All comparisons above are added to give a sum-total for each of the computer's generated 2²⁶muscle-sets.
  When this set of rules is applied to the example above it will change as follows:


Px:	−1 −1 0 0 1 1 2 2 3 3 −1 −1 0 0 1 1 2 2 3 3 −1 −1 0 0 1 1 2
mx:	0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0
Diff:	0 0 −2 2 1 −1 2 −2 3 −3 0 0 −2 2 1 −1 2 −2 3 −3 0 0 −2 2 1 −1 2

Sum − total = 2

- g. This method's sum-total values can range from 27*3=81 (when each Px=3 and each mx=0) to 27*(−3)=−81 (when each Px=3 and each mx=1). The muscle-sets smallest sum-total are considered best fit and are converted back to the nerve-sets that could have generated them (note that there may be more than one such nerve-set).
- h. From the latter, the one or the few nerve-sets with the smallest number of damaged nerves are output as the diagnoses.

The rules of comparison for sum-of-squared differences are identical to the above only that the sum-total is the sum of the squared-differences—however, each squared difference retains its original sign (!); in the example above the sum-total=4.

The logic behind this method is as follows: The aim is to build a system tuned to the severity of the abnormalities as stated by the clinician and to include those occurrences when Px>1 and mx=1 and those occurrences where Px=0 and mx=0. Those occurrences have no effect whenversion 1 of the rules of computation is applied (FIG. 4, A); the current rules will result in the inclusion of all the clinician data and levels of abnormalities. It is expected that these rules will enhance the success of the EMG ASSISTANT as a diagnostic system.

In order to implement said logic, agreement between Px and mx will generate negative values and will reduce the sum-total for each said comparison above, while a disagreement will generate a positive value that will increase the sum-total of each said comparisons above, and the best fit between the clinician's data and the computer generated virtual muscle-sets is the smallest sum.

Steps in the process relating to the lower extremities (LES):

These steps are identical to those of the UES in the confines ofFIG. 2: 27 muscles and 17 possible focal nerve-damage locations, amounting to only 2¹⁷(131,072) unique nerve-set combinations. Otherwise, the approach and computations are identical and will not be repeated here.

Additional Techniques

There is one variation to the above mentioned technique of generating and analyzing a 26×27 UES matrix (26 nerves by 27 muscles) or the 17×27 LES matrix (17 nerves by 27 muscles. This new technique accelerates the computational speed of the EMG ASSISTANT by many folds and gives the same diagnostic results, and can even accelerate matrices that are even much larger than 26×27, if more nerves and more muscles need be studied. First brief rehashing of the original technique as presented above inFIGS. 3, A to D: The program actually generates all the 2²⁶virtual and unique nerve-sets and converts each one to muscle-sets which are then compared one muscle at a time with the recorded clinician's muscle-set, the best-fits are converted back into nerve-set(s), and the diagnosis is the nerve-set(s) with the smallest number of damaged nerves. All this is done on the fly in real time and is lengthy. However, even though there are 2²⁶unique nerve-sets, there is not necessarily the same number of unique muscle-sets as the anatomical wirings of nerves-to-muscles force a smaller number of unique muscle-sets. This issue was studied and it was found that there can be only 5099 unique muscle-sets in the UES as presented in Table 1. The latter finding means that some muscle-sets can be generated by multiple combinations of unique nerve-sets.

Therefore, another computational path was designed to take advantage of the above, as follows:FIG. 3, A & B are computed as before only once, before the clinical study. Then the matrix inFIG. 3, B is searched for repetitions of the same muscle-set. When found, the nerve-set corresponding to it is kept in another matrix which is headed by that muscle-set. Thereafter, all repetitions are erased. We are left with 5099 unique muscle-sets and their varied number of unique nerve-sets that are possible diagnoses. As a consequence, the actual process of making a diagnosis, comparable toFIG. 3, D involve comparing—applying the various comparison-techniques above—only 5099 muscle-sets to the clinician's empirical set and when the best fit is found it is already coupled with the possible diagnoses. The last step is just like similar steps described before. Among those coupled diagnoses the one or several with the smallest number of damaged-nerves are chosen and output as the final diagnoses.

Output

Once a diagnosis is made, i.e. nerve-set(s) comprised of 26 nerves or nerve-segment names, some are marked as normal and some as damaged is output and the following statistics are computed and output with it:

- a. The damaged nerves are marked as in the example culminating inFIG. 5, A & B. There are two possible diagnostic nerve-sets (combination #1 and combination #2). In each, only one nerve was found damaged and marked with an arrow. Incombination #1, it is the lower trunk; while incombination #2 it is the C8 Root.
- b. Each diagnostic nerve-set is converted to a muscle-set usingFIG. 1 or 2, respectively.
- This muscle-set is called “Best-fit”. Ideally, it will be identical to the clinician's muscle-set, named “Input”. The closer they are the more reliable the diagnosis is.
- c. In each of the combinations (#1 and #2), separately, these two muscle-sets are numerically compared to each other to establish how close to the ideal fit they are. These numerical comparisons are as follows:
- d. A truth table summarizing the quality of said fit into four groups: True positives, false positives, true negatives and false negatives.
- e. A 2-sided exact probability Binomial Test p-value is computed comparing the recorded muscle-set and the best fit muscle-set. This statistic takes into account the total number of muscles that were studied, the number of muscles correctly classified (true-positives and true-negatives), and the number of muscles incorrectly classified (false-positives and false-negatives). The p-value indicates the likelihood that the conversion muscle-set best fit in the table and the original recorded muscle-set relate to each other better than just by chance.
- f. The percentage of correctly classified muscles is computed as 100*[(true-positives+true-negatives)/(total number of muscles examined)].

Step 5: Computing the next-best-muscle to sample.

The clinician may be unhappy with the results, such as when the resulting diagnosis is given an unsatisfactory p-value, or when the computations end with too many possible diagnoses (combinations) which is not very helpful whether statistically significant or not, or if for whatever reason the clinician wants to sample additional muscles. In such cases the EMG ASSISTANT can compute and advise the clinician which is the next-best-muscle to sample.

This computational decision is done as follows: The EMG ASSISTANT takes the muscle-set that was sampled by the clinician and adds to it one muscle from those muscles that were not sampled, once as if the added muscle is abnormal, for example +2, and once as if the added muscle is normal (0). Then it computes and keeps the usual statistics as inFIG. 5. Once the computations for the added muscle are completed, said muscle is dropped and a different muscle that previously was not recorded by the clinician is added, and the process repeats. At the end of this process the muscle that has the best statistics (see below) both when presented as abnormal and when presented as normal is the one suggested to the clinician as the next-best-muscle to sample. There may be more than one. There may be none, in which case the program reports in two columns, one column for the next-best-muscle or muscles that generate(s) the best statistics when abnormal and the other column for the next-best-muscle or muscles that generate(s) the best statistics when normal. The clinician is advised to take one muscle from each column to sample.

Said statistics include:

- 1. The binomial-test p-value, subsequently abbreviated as ‘p’,
- 2. The number of possible correct diagnoses, also named combinations inFIG. 5, subsequently be abbreviated as ‘nComb’, and
- 3. The percentage of correctly fitted muscles, subsequently abbreviated as ‘%’.

The actual decision making process goes as follows: Said muscles that are added and dropped are assumed abnormal,grade +2. For each one of said muscles the three statistics ‘p’, ‘nComb’ and ‘%’ are computed. These statistics are entered in one list.

Then, said muscles that are added and dropped are assumed to be normal,grade 0. For each one of said muscles the three statistics ‘p’, ‘nComb’ and ‘%’ are computed. These statistics are entered into another list

- a. Thereafter, the program will seek those muscles for which ‘p’ & ‘nComb’ & ‘%’ have the best values in both lists and report them as the next-best-muscle (or muscles) to sample. If there are more than one the clinician will choose his preferred one.
- b. If the criteria in (a) above could not be met, then the program will screen said muscles statistics for muscles with smallest ‘nComb’ and then, among the chosen ones, the program will select again those with the best ‘p’ & ‘%’ and report them as the next-best-muscle (or muscles) to sample.
- c. If none of the muscles shows on both lists, the clinician is expected to choose one best muscle from each list and sample them both.

The next-best-muscle suggestion is reported inFIG. 5,B last paragraph.